Unfiltered AI Models for Ollama vs a Hosted API
If you are looking at unfiltered ai models for ollama, the real question is where the model should run: on your own GPU or on someone else's. Local wins on privacy and offline use, hosted wins on setup, model size and no hardware bill.
Updated


Key points
- Local Ollama needs enough memory for the whole model; hosted needs only a key.
- Ollama and this API both speak the OpenAI format, so switching is a base URL change.
- Hosted cost here is $0.15 per 1M input and $0.50 per 1M output tokens, so light use costs cents.
- Choose local for offline or air-gapped work, hosted for always-on apps and long context.
Quick decision table
| Situation | Better fit |
|---|---|
| No GPU, or a laptop with little memory | Hosted API |
| Data must never leave your machine | Local Ollama |
| An app or bot that must be up around the clock | Hosted API |
| Offline or air-gapped use | Local Ollama |
| Long documents (up to 100,000 tokens of context) | Hosted API, unless your GPU has memory to spare |
| Heavy continuous generation on hardware you already own | Local Ollama |
| Trying a model for an afternoon | Either; hosted is faster to start |
Most people end up using both. The rest of the article explains why each row lands where it does, so you can adjust the table for your own case.
What running locally really takes
Ollama downloads a model file and runs it on your CPU or GPU. The first constraint is memory. A rough rule: the weights need about the number of parameters times the bytes per weight. At 4-bit quantization that is around half a byte per parameter, so a 7B model needs roughly 3.5 to 4 GB for weights, a 13B model about 7 GB, and a 70B model about 35 GB, before the cache that grows with context length.
Those figures are estimates, and quantization costs some answer quality. When the model does not fit in GPU memory, Ollama spills layers to the CPU and speed drops sharply. Long prompts make this worse, because the attention cache grows with every token you feed in.
There is also the model-choice problem. Community fine-tunes labelled unfiltered or abliterated vary a lot in quality, and each comes with its own prompt template quirks. Part of the cost of local use is your time spent finding one that behaves.
What a hosted endpoint changes
The hosted unfiltered AI model on this site is one open-weight model run by an independent provider. You do not choose quantization, manage drivers or wait for downloads. You get a key and call it.
- Context: 100,000 tokens, prompt plus completion.
- Features: streaming, tool calling, JSON mode, and the usual sampling parameters.
- Concurrency: up to 8 simultaneous requests on a full key, which a single home GPU often cannot match at the same quality.
- Privacy: prompts are not used for training, but they do travel to a server. Local is stricter on that point and nothing here changes that.
The trade is plain. You give up the guarantee that nothing leaves your machine, and you gain a model that is running when you need it, with no hardware to buy or cool.
Cost: when does buying hardware pay off
Hosted pricing is $0.15 per 1M input tokens and $0.50 per 1M output tokens, prepaid, with no subscription. For a chat turn of 800 input and 400 output tokens, that is $0.00012 + $0.0002 = $0.00032.
- 1,000 turns: $0.32
- 100,000 turns: $32
- 1,000,000 turns: $320
Now compare with hardware. Suppose, as a pure example, you spend $600 on a GPU just for this. Divide by $0.00032 and you get 1,875,000 turns before the hardware equals the hosted bill, and that ignores electricity and your time. If you already own a capable GPU, local use is free at the margin and the comparison flips.
The practical reading: hobby and moderate use costs a few dollars a month hosted, which rarely justifies a purchase. A pipeline generating tens of millions of tokens daily is another matter, and the sums deserve a proper spreadsheet. Current rates are always on the pricing page. For your own estimate, multiply your expected daily tokens by the two rates above.
Speed and reliability
Local speed depends entirely on your hardware and the model size. A model that fits in GPU memory feels instant; one that spills into system memory can feel painfully slow. This article quotes no token-per-second figures, because they vary too much between cards and quantizations to be useful.
Hosted speed depends on load. When the service is briefly overloaded it returns HTTP 429 with a Retry-After: 5 header, and your client should wait and retry. Over 300 requests per minute on a full key returns rate_limited, and more than 8 at once returns concurrency. Neither is billed.
The reliability difference is mostly about who does the maintenance. Your own machine needs updates, enough free disk space and an awake process. The hosted API needs a key with credit on it.
Switching tools between Ollama and the API
Ollama exposes an OpenAI-compatible endpoint on port 11434, and this API follows the same format. That means one script can talk to either, chosen by environment variables.
import os
from openai import OpenAI
if os.environ.get("USE_LOCAL") == "1":
client = OpenAI(base_url="http://localhost:11434/v1", api_key="ollama")
model = os.environ["LOCAL_MODEL"] # the name you pulled in Ollama
else:
client = OpenAI(
base_url="https://api.unfilteredaimodels.com/v1",
api_key=os.environ["API_KEY"],
)
model = "uncensored"
r = client.chat.completions.create(
model=model,
messages=[{"role": "user", "content": "Describe a rainy harbour town in two sentences."}],
)
print(r.choices[0].message.content)
The same trick works in chat front ends and agent tools that ask for an OpenAI-compatible base URL. Keep two profiles, one pointing at localhost and one at https://api.unfilteredaimodels.com/v1, and flip between them. The docs list the supported parameters if a tool sends something unusual.
Prompts you tuned locally usually carry over, but check them. A different model can respond to the same system prompt differently, so run a few of your own test prompts after switching.
Getting started with the hosted side
Create an account and visit the API key page. New users can claim $0.50 of trial credit with one click, valid for 7 days and needing no card. At 800 input and 400 output tokens per turn, that covers around 1,560 turns, enough to decide whether the hosted model suits your project before you spend anything.
Paid top-ups start at $20 with any whole amount up to $500, and credit never expires. If you keep an Ollama setup as well, you lose nothing: use local for private drafts and the hosted API for the jobs that need uptime, long context or concurrency.
More guides
Questions and answers
Your key is one form away
Create an account, copy the key, change the base URL. That is the whole setup.
Can Ollama run unfiltered models?
- Yes, Ollama runs whatever compatible model files you pull, including community fine-tunes that refuse less. Quality and prompt format vary by model, so test before relying on one.
Is a hosted API private enough?
- Prompts are not used for training, but they are processed on a remote server. If nothing may leave your machine, local is the stricter choice.
Can I use my Ollama client code with the hosted API?
- Often yes. Both use the OpenAI format, so you change the base URL to https://api.unfilteredaimodels.com/v1, add your key and set the model to uncensored.
How much does the hosted option cost per month?
- It depends on volume. At $0.15 per 1M input and $0.50 per 1M output tokens, 100,000 chat turns of 800 input and 400 output tokens cost about $32.